Here is our paper on learning a program for physical reasoning that generalizes to variable numbers of objects and different scene configurations. It would be great to start a discussion, get feedback, and explore ideas for future work. I've linked the project webpage and the paper.